A Human Behavior Recognition Method Based on the Fusion of Multi-feature Spatial-Temporal Relationships

Abstract

The invention discloses a human behavior recognition method based on the fusion of multi-feature spatio-temporal relations. The specific steps include: expressing the dense track features extracted from the video through the optical flow histogram and the motion boundary histogram, and then using the KMEANS algorithm to construct the two features Corresponding to the spatio-temporal bipartite graph between the centroids, the K-way bipartite graph segmentation technology is used to segment the spatio-temporal bipartite graph, and the representation method based on conditional probability is used to obtain the video-level code after the fusion of the two features, and finally the classifier is trained and recognized , through the above method, the present invention is a human behavior recognition method based on multi-feature spatio-temporal relationship fusion, the method calculates the spatio-temporal distance between the features in each video, thereby constructing the spatio-temporal bipartite graph between the centroids corresponding to the two features , the K-way bipartite graph segmentation technology is used to segment the spatio-temporal bipartite graph, and the centroid with strong spatio-temporal relationship is fused, which better excavates the effective information of different features and improves the recognition accuracy.

Description

technical field [0001] The invention relates to the field of computer vision, in particular to a human behavior recognition method based on multi-feature spatio-temporal relationship fusion. Background technique [0002] With the development of computer science, video has become a part of people's lives. How to make computers "understand" human behavior in video plays an important role in the fields of content-based video retrieval, intelligent monitoring, human-computer interaction and virtual reality. [0003] Generally speaking, a classic human action recognition framework mainly includes three steps: feature extraction, video coding, and classifier training and recognition. In addition, for the case of using multiple features, it also includes an optional multi-feature pre-fusion Or post-fusion step, where video encoding is a key step in determining the accuracy of recognition. [0004] At present, one of the widely used and improved encoding methods is the Bag of Words...

AI Technical Summary

Problems solved by technology

[0004] At present, one of the widely used and improved encoding methods is the Bag of Words (BoW for short) method. The classic BoW method first clusters the features, and then expresses the video as a frequency histogram of the features appearing in each centroid. Vector, although BoW coding has shown good generalization ability and robustness in many literatures, but this method also has many shortcomings: such as time-consuming feature clustering process, KMEANS algorithm supervised parameter k and the spatio-temporal relationship between centroids loss of information

Due to the need to construct multi-level centroids, the computational complexity of this method is relatively high

Images